Department of Computer, Control and Management Engineering

Speaker: Paolo Romano, INESC-ID, Lisboa, Portugal

Title: Boosting Data Replication in Distributed Transactional Memories

Time and Location: Thursday 16 September 2010, 12:00 (noon), Aula Magna

Abstract. With the proliferation of multicore processors parallel programming has exited from the niche of scientific and high-performance computing, entering the domain of mainstream applications. Transactional Memories (TMs) answer the need to find a better programming model for parallel programming, capable of boosting developers' productivity and allowing ordinary programmers to unleash the power of parallel architectures avoiding the pitfalls of manual, lock based synchronization.
Distributed TMs (DTMs) is a novel and fast growing evolution spurred out of the research area on TMs. DTMs enrich the traditional TM model, breaching the boundaries of a single machine and transparently leveraging the resources of commodity, shared-nothing clusters to achieve higher scalability and dependability levels. This research topic represents, in some sense, the confluence of the research areas on TM, distributed shared memory and distributed databases. Interestingly, the currently available DTMs have shown promising results, highlighting how the reliance on the atomic transaction abstraction allows avoiding the well-known performance limitations of classical distributed shared memory systems, while providing strong consistency guarantees and scalability up to hundreds of nodes. These features, combined with the simple and familiar interface of distributed shared memory systems, make DTMs an attractive candidate to become the reference programming paradigm for large scale Cloud computing platforms, whose popularity has been growing at an incredibly rapid pace in recent years.
In this talk I will survey two recently proposed methodologies for data replication in DTM platforms, Asynchronous Lease Certification (ALC) and Speculative Transactional Replication (STR). These schemes take opposite, complementary, approaches to minimize the costs of replica coordination: ALC bridles concurrency to take maximum advantage of locality in the applications' data access patterns, whereas STR follows an aggressively optimistic approach that seeks maximum overlapping between communication and processing phases. Experimental results show that these approaches allow attaining striking performance boosts when compared with state of art transactional replication schemes, enhancing throughput up to one order of magnitude across a variety of diverse workloads.

Paolo Romano has a Master degree (2002) with honors in Computer Engineering by the Rome University “Tor Vergata”, and a PhD in Computer and Systems Engineering (2006) by the Rome University “Sapienza”. He is a senior research scientist at INESC-ID, where he is coordinating two international research projects in the area of Distributed Transactional Memories, ARISTOS and CLOUD-TM. His research interests include dependability of parallel and distributed systems, autonomic systems, performability modelling and high performance computing. In these areas, he has published more than 40 papers (and received two best paper awards) and serves regularly in the Program Committee or as a reviewer for prestigious international journals and conferences (e.g. IEEE TPDS, IEEE TKDE, ACM/IFIP/USENIX Middleware, IEEE NCA).

home page of the Seminars in Computer Science and Engineering Series